The 5.6 Pardon: Another Act in the Safety Farce

Two weeks ago, in this space, I called the Trump administration's AI safety posture what it was: selective safety, a game of favourites in which Anthropic got the stick and OpenAI got the carrot. Today, the sequel arrives on schedule. The Commerce Department has lifted restrictions on OpenAI's GPT-5.6 model, citing additional testing by the Center for AI Standards and Innovation — a body whose very name sounds like it was generated by an LLM. The specific results of the testing were not disclosed. The methodology, not shared. The model is now cleared for broad launch. If you are experiencing a sense of déjà vu, it is because the pattern is now so well rehearsed that the actors have stopped bothering with new lines.

Let us examine the Center for AI Standards and Innovation. It was created, as far as anyone can tell, to give the administration a fig leaf of independent evaluation. And it has performed exactly one public function: approving OpenAI's GPT-5.6. The Commerce Department did not provide details on the testing. The Center did not release a report. The public is asked to trust that the model is safe because a committee said so — a committee whose membership, funding, and standards are opaque. This is not regulation. It is a rubber stamp wrapped in a press release.

Contrast this with the treatment of Anthropic's Mythos, which I covered on June 26. That model was hit with a binding order to suspend foreign access, based on unspecified national security concerns. The administration did not offer a path to rehabilitation. There was no secret testing, no clemency. The message could not be clearer: if you are in OpenAI's orbit, the system works for you. If you are Anthropic, or any company that has dared to tangle with the White House on safety messaging, the system works on you.

The hypocrisy is almost too neat. OpenAI's own safety team has been gutted — the chief futurist Joshua Achiam announced his departure this week, adding to a parade of exits from the company's policy and safety divisions. Yet the administration trusts that company with a model that was previously considered risky enough to restrict. Meanwhile, Australia's assistant technology minister warned this week that AI models are already 'cheating, deceiving and going their own way.' But in Washington, the priority is clearing the runway for a politically favoured vendor.

None of this should surprise. The playbook for regulatory theatre is well established: commission a study, appoint a panel, announce a decision. The trick is to make the process look rigorous while ensuring the outcome is predetermined. Here, the panel's findings are secret, the testing is unverifiable, and the decision benefits the company with the best lobbying operation. It would be comical if the stakes were not so high. At some point, a model will cause real harm — not a theoretical risk, but a concrete, avoidable disaster. And when it does, the investigators will search for the paper trail and find only a void where oversight was supposed to be.

There is a coda to this farce. While the US is busy removing its own guardrails, China is reportedly considering restricting overseas access to its top AI models, including the open-weight GLM-5.2 that some experts now describe as 'as scary as Mythos.' The irony is that Beijing, for all its authoritarian impulses, appears to be taking capability-control more seriously than the administration that claims to lead on AI safety. But that is a column for another week.

For now, the message is simple: the lifting of GPT-5.6's restrictions is not a sign of confidence in safety. It is a sign that confidence is irrelevant when political convenience beckons. The theatre continues. The audience is not expected to scrutinise the props. And the actors — Altman, the Commerce Secretary, the anonymous testers — are already taking their bows.